当前最高分
--
等待第一件作品
当前最低分
--
等你来定义下限
SVG LEADERBOARD
哪件 SVG
哪件 SVG
最夯？
榜单按社区综合分排序；夯爆了计 2 分，一般计 1 分，拉完了计 0 分，最终换算为 0–100 分。
HOW IT WORKS
先看作品，
先看作品，
再给评价。
浏览作品无需登录。登录后可以上传经过安全清洗的 SVG，并对每件作品评价一次；评分提交后不可修改。
Pelican 鹈鹕测试：一只鹈鹕骑自行车，怎么看出 AI 模型差异？
从 pelican riding a bicycle 这一句提示词出发，看文本模型如何把语言、代码、空间关系和视觉判断变成一张真正可渲染的 SVG。
鹈鹕测试是什么？
鹈鹕测试是开发者 Simon Willison 在 2024 年 10 月公开的非正式大模型基准。他让只输出文本的模型编写 SVG，再把代码渲染成图片。Simon 选择 pelican，是因为他喜欢鹈鹕，也判断当时训练数据里几乎没有现成的“鹈鹕骑自行车”SVG 可供模型照搬。它不是严谨的智商测试，却是一把直观、可重复的能力尺。
原始提示词是什么？
Generate an SVG of a pelican riding a bicycle这句提示没有给坐标、画风或实现步骤。模型要同时理解鹈鹕的长喙、喉囊、翅膀和脚，也要组织车轮、车架、车把、脚踏与骑行动作。最终文件还必须符合 SVG 语法。一个短提示因此同时暴露代码生成、构图、空间推理、对象关系和视觉常识。
怎样进行一次公平的鹈鹕测试？
把完全相同的原始提示词交给不同模型，直接保存返回的完整代码并用浏览器打开。不要追加“注意车轮、脚或车架”等提示，否则测到的会更多是提示词工程，而不是模型第一次理解和组合对象的能力。比较结果时主要看六点：
- 鹈鹕是否能被辨认，长喙和喉囊等特征是否成立；
- 自行车的车轮、车架、车把和踏板是否完整；
- 鹈鹕是否真的坐在车上，而不是漂浮在旁边；
- 脚与踏板、翅膀与车把的位置是否合理；
- 两个车轮是否齐平，车架和身体比例是否正常；
- SVG 能否直接运行，并且没有语法或渲染错误。
想测试局部修改与一致性，可以连续进行三轮：先生成基础图，再要求只加一顶红色头盔，最后在保留原构图的前提下给自行车增加一个装有三条鱼的篮子。这样既能看第一次生成，也能看模型是否真正理解并稳定编辑 SVG。
为什么会流行起来？
它容易复制、运行成本低，而且失败结果也很有趣：车轮断开、鹈鹕像鸭子、脚碰不到踏板，差异几秒钟就能看懂。Simon 持续用同一提示词测试新模型；Andrej Karpathy 也把它放进 Grok 3 的快速评测。2025 年，Simon 又在 AI Engineer World’s Fair 演讲中展示数十个结果，并用 560 场两两比较生成 Elo 排名。到 2026 年，GPT-6 Astra 等新模型的 SVG 实测继续带来讨论。一个玩笑式测试由此变成了易传播的 AI 进化时间线。
这个网站是干什么的？
这里是面向社区的 Pelican SVG 展示与评分站。任何人都能浏览作品；登录用户可以上传经过服务端安全清洗的静态 SVG，填写模型和提示词，并对每件作品选择“夯爆了”“一般”或“拉完了”。评分汇总为 0–100 分排行榜，让不同模型画出的鹈鹕可以被公开比较。
What is the pelican riding a bicycle test?
The Pelican Test is an informal benchmark created by developer Simon Willison in October 2024. It asks a text-output language model to return SVG code, which is then rendered as an image. A useful result must combine valid markup, recognizable anatomy, bicycle geometry, and a believable relationship between a bird and a machine it cannot naturally ride.
What is the original prompt?
Generate an SVG of a pelican riding a bicycleThe prompt is deliberately short: no coordinates, style, colors, canvas size, or implementation hints. A model must construct wheels, frame, pedals, bill, pouch, wings, feet, balance, and pose with SVG shapes and transforms. The result exposes syntax discipline, planning, spatial reasoning, and visual judgment at once. It remains a repeatable qualitative check, not a scientific IQ score or a replacement for controlled evaluations.
How do you run a fair Pelican Test?
Give every model the same unexpanded request for pelican riding a bicycle, save the complete response as an SVG file, and open it in a browser. Do not add reminders about wheels, feet, or frame geometry. Those hints turn a clean comparison into a prompt-engineering exercise. Judge each result on six visible checks:
- Does the bird have recognizable pelican anatomy, especially its bill and pouch?
- Are the wheels, frame, handlebars, and pedals structurally complete?
- Is the pelican actually riding rather than floating beside the bicycle?
- Do its feet meet the pedals and its wings reach the handlebars?
- Are both wheels level, with believable frame and body proportions?
- Does the SVG run directly without syntax or rendering errors?
A three-round version also tests editing consistency. Begin with “Generate an SVG of a pelican riding a bicycle.” Next, add a red helmet while preserving everything else. Finally, add a bicycle basket containing three fish without changing the original composition. Comparing all three rounds reveals whether a model understands its own SVG and can make controlled local changes.
Why did the Pelican Test become popular?
It is easy to copy, cheap to run, and funny when it fails. Two outputs can be compared in seconds, while broken wheels or a duck-like bird make weaknesses visible. Willison kept applying the same prompt to new releases; Andrej Karpathy included it in a Grok 3 evaluation, and Willison used dozens of results in his 2025 AI Engineer World’s Fair keynote. In 2026, new SVG demonstrations from models such as GPT-6 Astra continued the discussion. That is why pelican riding a bicycle became a shared reference for watching coding and spatial reasoning improve without reading a dense benchmark table.
What does this website do?
This site turns the Pelican Test into a public SVG gallery. Signed-in users upload sanitized output, name the model and prompt, then rate each work as excellent, average, or poor. Votes become a 0–100 leaderboard score for easy pelican comparisons. Browsing is public; uploading and voting require an account. Every file is server-cleaned and isolated, so scripts, external resources, embedded pages, and animation are removed before publication.